UA-UKAN: a unified attention-enhanced Kolmogorov-Arnold network for medical image segmentation
TL;DR
Experimental results indicate that the proposed network effectively improves noise robustness and irregular lesion boundary reconstruction accuracy, achieving stable and high-precision segmentation across multimodal medical imaging tasks.
Abstract
To address the issues of shallow texture loss, weak noise robustness, and the difficulty of jointly modeling global context and local lesion boundaries in existing medical segmentation networks, this paper proposes a UA-UKAN segmentation network. In the encoder, UCAN is designed to hierarchically enhance fine-grained features and adaptively suppress noise, while an improved Token-KAN is introduced to recover spatial information lost during tokenization. In the decoder, GLTransKAN integrates global-local attention with KAN nonlinear transformation to achieve simultaneous global constraint and lesion boundary reconstruction. Extensive experiments are conducted on three multimodal medical image datasets, including BUSI, CVC-Clinic DB, and ISIC-2017. Quantitative results demonstrate that the proposed method consistently outperforms mainstream baselines. Ablation studies verify that UCAN and GLTransKAN provide independent improvements and complementary benefits. Visualization results further show that UA-UKAN achieves more accurate lesion contour fitting and stronger interference resistance. Experimental results indicate that the proposed network effectively improves noise robustness and irregular lesion boundary reconstruction accuracy, achieving stable and high-precision segmentation across multimodal medical imaging tasks.